In my last post, I discussed two important characteristics of delta-sigma analog-to-digital converters (ADCs) that simplify the design of your anti-aliasing filter: an oversampling architecture and a supplemental digital decimation filter. The oversampling architecture places the Nyquist frequency further away from your signal bandwidth of interest, while the digital decimation filter attenuates most of the unwanted out-of-band signals. When combined, they allow for a more relaxed anti-aliasing filter response, which you can achieve with only a few discrete components.

**Figure 1. Keep those aliases back with a proper anti-aliasing filter**** **

We know that using anti-aliasing filters is beneficial in precision ADC applications, but designing the right anti-aliasing filter is just as important – if you’re not careful, you can introduce unwanted errors into your system just as easily as you can remove them. Consider these three general guidelines when designing the anti-aliasing filter for your application.

**Choose your filter cutoff frequency**

The simplest anti-aliasing filter is a single-pole, low-pass filter using a series resistor (R) and a common-mode capacitor (C_{CM}), as shown in Figure 2. The first step in designing this filter is to select the desired cutoff frequency, f_{C}. At f_{C}, the filter’s response rolls-off to -3dB and continues to decrease at -20dB/decade in the frequency domain.

Choose a cutoff frequency at least a decade less than the ADC modulator sampling frequency, f_{MOD}, in order to knock down the out-of-band noise at those frequencies by a factor of 10 or more. For increased attenuation, reduce the cutoff frequency further by increasing the values of R and C_{CM}. Remember from my last post, your digital decimation filter is there to help too, so it’s not necessary to set your anti-aliasing filter cutoff immediately after your signal bandwidth of interest.

Equation 1 calculates the -3dB cutoff frequency for a single-pole, low-pass filter:

** **

** **

**Figure 2. Single-pole, low-pass filter at the ADC inputs**

Occasionally, a single-pole, low-pass filter may not be enough. Some applications, like vibration sensing, may use less oversampling to analyze signals over a wider bandwidth. This puts the passband of the digital decimation filter closer to f_{MOD} and gives the anti-aliasing filter less room to roll-off. In these instances, you can add a second or third pole with additional RC pairs to achieve an even sharper filter response.

Figure 3 illustrates the response of a single-pole and a double-pole filter designed for an ADC, sampling the inputs at f_{MOD} = 1MHz. The double-pole filter allows the flat passband to extend out to about 20kHz and still achieve -60dB of attenuation at 1MHz.

**Figure 3. Frequency response of a single-pole and double-pole low-pass filter**** **

**2. Consider differential vs. common-mode filters**

Many ADCs convert the differential voltage between two independent inputs (i.e. INP and INN), so designers often include common-mode filters on each input to maintain system common-mode rejection (CMR). However, component tolerances will introduce mismatch between any two filters and degrade CMR across frequency as they filter common signals differently. This produces a differential signal error through what is known as common-mode-to-differential conversion.

Equation 2 calculates the CMR of your common-mode anti-aliasing filters at a given frequency, f, using the resistor tolerance, R_{TOL}, and the capacitor tolerance, C_{TOL}:

For applications that require high CMR, consider adding a differential filter to supplement the two common-mode filters as shown in Figure 4. Set the differential cutoff to be one decade less than the common-mode cutoff by increasing the differential capacitor, C_{DIFF}, by a factor of 10 over C_{CM}. This will help mitigate errors introduced by common-mode component mismatch and create a sharper overall filter response. Equation 3 calculates the cutoff frequency for a differential low-pass filter. Notice the extra factor of 2 in the denominator.

** **

** **

**Figure 4. Common-mode filters with the addition of a differential filter**** **

**3. Select appropriate component values**

Adding resistors to the signal path will introduce unwanted noise and errors in the measurement, so it is important to keep them reasonably small whenever possible.

Resistor noise – also known as Johnson or thermal noise – can be modeled as a voltage source in series with your ideally “noiseless” resistor. In general, you don’t want the resistor thermal noise to dominate the signal chain, so it is important to keep it below the noise floor of your ADC. Equation 4 calculates the noise density, v_{n}, of resistor thermal noise:

where k = Boltzmann’s constant (1.38E-23 J/K) and T is the temperature in degrees Kelvin.

Series resistors also introduce a small offset voltage in the presence of input bias currents. While you may be able to calibrate this later, limit the size of your resistors as much as possible, especially when bias currents are expected to be large.

Unlike the filter resistors, the higher the capacitance you can use, the better. The reasoning behind this has to do with how ADCs sample the inputs.

The inputs of a delta-sigma ADC without an integrated buffer are connected directly to the switched-capacitor sampling structure of the ADC modulator. This sampling structure is comprised of a network of switches and sampling capacitors on the order of 10pF or 20pF. Figure 5 is a simplified example.

**Figure 5. Simplified switched-capacitor sampling structure in an ADC**** **** **

The switched-capacitor circuit places a transient load on the external circuitry during sampling. The filter capacitors help attenuate the sampling charge injection from the modulator and provide some of the instantaneous current needed to charge the sampling capacitor, C_{SAMPLE}. The larger the filter capacitors, the more charge that will be available. Use NP0/C0G type ceramic capacitors because of their high Q-factor, low-temperature coefficient, and stable electrical characteristics. Larger capacitance values may also improve AC specifications like total harmonic distortion (THD), but keep in mind that this increases the filter’s RC time constant and requires a longer settling time.

I hope these three guidelines have prepared you to go about designing your next anti-aliasing filter. More importantly, I also hope that this blog series has taught you how aliasing occurs in data-acquisition systems and where delta-sigma ADCs hold an advantage over other ADC architectures. If there are other points that I did not cover, feel free to leave a question or comment below!

**Additional resources**

- Check out other posts by my colleagues with design tips and other delta-sigma ADC basics, including one I wrote on aliasing in ADCs and the advantages of the delta-sigma architecture.
- Find more than 100 data converter technical resources in our Data Converter Learning Center.

Just a thought about single-pole filters, and 'flat':

An RC filter is never 'flat', it has roll-off beginning at any finite frequency, correct?

It looks like the amplitude error of a single-pole filter, let's say 1 kHz 3 dB point, at 100 Hz seems to have about 1% attenuation, and at 10 Hz, it's still .01% ... therefore, it seems that we can have an error of 1 LSB at only 13 bits resolution, at 10 Hz... to get useful 16-bit resolution, the bandwidth is down to a couple of Hz.

I think that we often use delta-sigma converters for high-resolution applications, so it seems that a single-pole filter could limit accuaracy at relatively low frequencies, relative to the 3 dB point? Is this reasoning at all correct?

Barry,

You are correct that there is roll-off below the cut-off frequency (-3dB). For RC low-pass, Vo/Vi = 1/SQRT(1 + (F/F0)^2). So, at 0.1xF0, the attenuation is 0.005 (0.5%); at 0.01xF0, it's 0.00005 (good for about 14+ bits); and at 0.0055xF0, it's good to 1LSB in 16 bits.

If you know the frequency of interest (single frequency, or a reasonably narrow FFT bin), you can compensate for the error by using a formula or table, e.g. at 0.1xF0, just divide by 0.995, etc.

Another source of roll-off error is the decimation filter (sinc3 filter typical), if its cut-off isn't much greater than the highest frequency of interest. This can be similarly compensated for.

Ryan,

Eq. 2 appears incorrect. As the tolerances become tighter, the first term will go more negative, reducing the calculated CMR. Obviously, tighter tolerances should allow CMR to approach its perfectly matched (theoretical) value, i.e. the first term should approach 0. Also, the second term goes negative for F < Fc, which doesn't make sense to me. Below Fc, CMR should approach 0, I would think.

I agree with Eq. 3, but the text preceding it says to make Cdiff = Ccm * 10 to make the differential-mode cutoff a decade lower than the common-mode cutoff. This factor for Cdiff would appear to make Fc(diff) = Fc(cm) / 20, which doesn't match the text. Of course, since the article started by selecting a cutoff frequency for anti-aliasing, I would suggest this as the desired differential-mode cutoff, and say that the common-mode cap values should be reduced by a factor of 10 so that the common-mode filter mismatch due to RC tolerances has little impact on the differential-mode filter response.

The conclusion I draw from reading some other articles about single-pole anti-aliasing filters for differential-input A/Ds, is that with perfectly matched components, one wouldn't need a differential-mode cap at all, and with no common-mode noise, one would need *just* the differential-mode cap. But in the real world, we need both: differential-mode cap for the "main" anti-aliasing function, and common-mode caps to filter as much common-mode noise as possible without significantly affecting the anti-aliasing filter response [due to common-mode to differential-mode conversion]. It would be interesting to study the effects of varying the RC tolerances and Ccm to Cdiff ratio to find an optimal solution (for a given noise environment).